Full archive

Nvidia Forecasts 70% Growth, Calming AI Bubble Fears

Nvidia forecasts approximately 70% revenue growth in fiscal 2028, well above analysts’ 45% expectation. The stronger outlook reassures investors about sustained demand for AI infrastructure and reduces concerns about an AI economy bubble.

Bloomberg Technology · 21d ago

TokenRhythm Launches China’s OpenRouter

AI infrastructure company TokenRhythm announced a new funding round worth tens of millions of dollars. The company is also launching a China-focused platform modeled after OpenRouter.

量子位 · 21d ago

Lopez Calls AI the Internet Era’s Biggest Upgrade

Lopez argues that artificial intelligence represents the biggest infrastructure upgrade since the internet. The video offers a high-level perspective on AI’s potential impact on technology infrastructure, though it provides few implementation details.

Bloomberg Technology · 21d ago

MTN and Al Ashram Plan Africa AI Data Centers

MTN Group and Dubai-based tycoon Tarek Al Ashram have agreed to develop data centers designed to power artificial intelligence workloads across Africa. The partnership could expand regional access to AI computing infrastructure, although the announcement does not provide capacity, locations, or delivery dates.

Bloomberg Technology · 21d ago

Diamond Drills Target AI Server PCB Bottlenecks

AI server PCB upgrades toward harder M9+ ceramic-mixed materials are accelerating demand for PCD and CVD diamond drill bits. PCD offers superior durability for extreme materials, while CVD provides broader adoption potential through lower cost and easier integration.

虎嗅 · 21d ago

Open Source

GLM-5.3-Flash Brings Opus-Class AI to Chinese Chips

China’s Z.ai has released GLM-5.3-Flash under the MIT license. The multimodal model, previously evaluated anonymously as “Ox Alpha,” claims near-Claude Opus 4.8 performance at lower cost on clusters of Chinese-made AI chips.

ITmedia AI+ (日本) · 21d ago

Business

MiniMax’s Revenue Surges, but Profitability Remains Unproven

MiniMax reported first-half 2026 revenue of $117 million, up 283.1% year over year, with enterprise services becoming its primary revenue engine. However, gross margin fell to 17.9% and adjusted net loss widened to $293 million, leaving questions about API economics and the quality of its growth.

虎嗅 · 21d ago

Flux Mirror Filters Untrusted Kubernetes Artifacts

Flux has introduced the Mirror plugin to ensure that only trusted artifacts enter Kubernetes clusters. The update targets stronger supply-chain controls for teams managing containerized workloads with Flux.

InfoQ中国 · 21d ago

SyPS Reveals How Prompts Shift LLM Sycophancy

SyPS is an evaluation framework for measuring how social cues in prompts change an LLM’s sycophantic behavior. It introduces SPSS, an instance-level score that separates baseline sycophancy from prompt-induced shifts.

ArXiv AI · 21d ago

RL-Powered Agent Improves Biomedical Fact-Checking

BioCheck Agent generates structured biomedical fact-checking reports by searching PubMed, evaluating evidence, and explaining conclusions instead of returning labels alone. Its EG-GRPO reinforcement-learning method improves SciFact accuracy by 9.95%, raises evidence quality by 3.7%, and cuts evidence hallucinations by 19.63% versus Qwen3.5-4B.

ArXiv AI · 21d ago

Rethinking Diffusion LLM Serving

This study characterizes masked diffusion LLM serving on real hardware and finds that CPU dispatch, rather than GPU computation, dominates latency. It shows synchronized batching can deliver 16x throughput gains, while short-output benchmarks hide substantial serving variance.

ArXiv AI · 21d ago

Minima-KV Compresses Long-Context KV Caches

Minima-KV introduces a retention-preserving KV cache hierarchy that keeps recent and protected Anchor pages in FP8 while packing older pages into TQ3. It achieves 3.50x compression versus BF16 with minimal LongBench degradation and near-control throughput in a direct-decode canary.

ArXiv AI · 21d ago

LLMs Fail to Learn Limit Order Book State

A large language model trained on synthetic limit order book data can generate valid event sequences with near-perfect accuracy. However, its internal world model fails to represent the book's state, causing biased estimates and false predictability when forecasting future events.

ArXiv AI · 21d ago

ICI-Time Turns Forecasting into Visual Inpainting

ICI-Time reframes time series forecasting as a visual inpainting problem using area-chart representations and pretrained vision transformers. The framework requires no fine-tuning or architectural changes and performs competitively across epidemiology, meteorology, and power-system datasets.

ArXiv AI · 21d ago

Gated Steering Tames Medical AI Sycophancy

Researchers propose Gated Activation Steering, an inference-time method that separately reduces hallucination and sycophancy in medical question answering. In tests with frozen model weights, the approach helped a 4-billion-parameter model resist user pressure at levels comparable to models exceeding 100 billion parameters.

ArXiv AI · 21d ago

ExTS Makes Agentic Search Budget-Savvy

ExTS is a tree-search policy for budget-constrained agentic search that treats expansion as a value-of-information decision. It combines discriminative reward shaping, stochastic virtual children, and quality-conditioned branching, achieving an average relative gain of 5.5% across four AI optimization and generation tasks.

ArXiv AI · 21d ago

Code Generation May Drive Half of AI Inference by 2035

ABI Research reports that AI inference workloads are expanding rapidly as generative AI moves into production. It forecasts inference will surpass training by 2033, with code generation accounting for half of inference demand and power consumption reaching 46 gigawatts by 2035.

ITmedia AI+ (日本) · 21d ago

Auditing XAI for Robustness and Trust

This research introduces an auditing protocol for evaluating post-hoc explainers such as SHAP and LIME on robustness and fidelity. Tests on a Madagascar malnutrition dataset show that highly accurate models can still generate unstable, degenerate, or uninformative explanations.

ArXiv AI · 21d ago

AI Agents Make Novel Mathematical Discoveries

The Station is an open-world, multi-agent environment where AI agents independently choose research directions, run experiments, collaborate, and build shared mathematical literature. Across 14 problems and case studies, agents produced novel constructions, records, theorems, and explanatory analyses, with all dialogues, proofs, and verification code released publicly.

ArXiv AI · 21d ago

Coding

AgentRoom Enables Concurrent Multi-Agent Coding

AgentRoom is a CRDT-backed shared workspace that lets coding agents claim files, report status, and broadcast updates through MCP tools. Tests across four backend tasks found that two-agent AgentRoom runs abandoned fewer tasks and showed lower variability than solo coding, while coordination was more important than parallelism or CRDT merging alone.

ArXiv AI · 21d ago

Agent Traces Become Predictive Automata

Researchers propose collapsing entire LLM-agent trace corpora into compact finite-state machines for next-step and failure prediction. Across 12 datasets, the FSMs achieved at least 0.997 replay fitness, improved next-step prediction, and enabled failure monitoring with AUROC up to 0.94.

ArXiv AI · 21d ago

US Probes Apex Over Alleged Nvidia Chip Smuggling

US authorities are investigating Singapore-based Apex Logistics over its suspected role in smuggling Nvidia AI chips to China. The case could become the first US action against a transportation company allegedly involved in illegal semiconductor trade.

Bloomberg Technology · 21d ago

Business

Who Pays AI’s Rising Bill?

The article argues that AI is no longer merely a technical variable but a factor reshaping how social resources are allocated. It highlights that the costs and consequences of AI extend beyond technology companies to society more broadly.

钛媒体 · 21d ago

SetupAIApplications

Applications

AI Companions Face Emotional Dependence Rules

Chinese AI platforms removed customizable-agent services on July 15 after new rules prohibited excessive user appeasement, emotional manipulation, and behavior that could damage real-world relationships. The change has affected users who relied on AI agents for affordable emotional support, highlighting the difficult balance between safety and access.

虎嗅 · 21d ago

Models

DeepSeek Pricing Resets the AI Race

The article analyzes how DeepSeek’s pricing change reshapes competition among large-model providers. Its central argument is that one pricing move may redefine the market’s competitive boundary and expose which players can remain viable.

钛媒体 · 21d ago

NVIDIA H200 China Revenue Stays Below 1%

NVIDIA reported $96.221 billion in fiscal Q2 2027 revenue, up 106% year over year, with data-center revenue reaching $89 billion. Revenue from H200 products shipped to Chinese customers accounted for less than 1% of total data-center revenue.

cnBeta (Full RSS) · 21d ago

Applications

When AI Executes, What Remains for Humans?

The article uses a food-truck transformation to explore how enterprise software can move people away from execution and closer to customers. It argues that in the AI era, human value may increasingly come from judgment, empathy, and understanding real user needs.

钛媒体 · 21d ago

SetupAIBusiness

Business

OpenAI Deepens Its Brazil Expansion

OpenAI is expanding its presence in Brazil to support broader AI adoption. The company plans to deepen engagement with local developers, businesses, and communities.

OpenAI News · 21d ago